🌟 Summary
The release of v8.3.37
introduces significant improvements and fixes across the export functionality and model operation modes, aiming to streamline user experience and enhance performance when using Ultralytics tools.
📊 Key Changes
- TensorRT Auto-Workspace Size: Implements an auto-managed workspace size for TensorRT exports by default, allowing more flexibility and reducing manual configuration errors.
- Label Padding Adjustment: Optimized the label augmentation by correctly updating vertical and horizontal padding, enhancing image annotation accuracy.
- Model Evaluation Mode: Introduced an
eval
method to easily switch models between training and evaluation modes, ensuring consistent performance during model assessments. - Documentation Updates: Added video tutorials for better understanding of hand keypoint estimation and annotation utilities, and standardized dataset configuration references for clarity.
🎯 Purpose & Impact
- Ease of Use: Setting the TensorRT
workspace
toNone
by default takes the burden off users to configure export parameters manually, simplifying the model export process. - Improved Accuracy: The fix in label padding ensures accurate annotations, critical for reliable model training and evaluation.
- Consistent Evaluation: By allowing models to switch to evaluation mode seamlessly, users will experience more reliable model performance metrics which are crucial for assessments.
- Enhanced Learning Resources: With new video tutorials, users can gain a deeper understanding of utilizing Ultralytics features, potentially increasing the adoption and correct usage of functionalities.
- Documentation Consistency: Transitioning to a uniform dataset configuration in examples reduces confusion, making it easier for users to follow guides and setups.
What's Changed
- Fix labels padding for Letterbox with
center=False
by @Y-T-G in #17728 - Add https://youtu.be/c-S5M36XWmg & https://youtu.be/fd6u1TW_AGY to docs by @RizwanMunawar in #17722
- Update
coco-seg.yaml
tococo.yaml
by @Y-T-G in #17739 - Bump astral-sh/setup-uv from 3 to 4 in /.github/workflows by @dependabot[bot] in #17753
- Standardize default region points in docs by @RizwanMunawar in #17721
- Enable model.eval() usage for
YOLO
class by @Laughing-q in #17754 ultralytics 8.3.37
TensorRT auto-workspace size by @Burhan-Q in #17748
Full Changelog: v8.3.36...v8.3.37